HVI-guided spatial-frequency fusion network for low-light image enhancement
TL;DR
The Spatial-Frequency Fusion Network (SFF-Net), which decouples intensity and chroma in HVI space and then learns frequency-aware enhancement with cross-branch compensation, is proposed, which improves SICE, Sony-Total-Dark, and unpaired real image results with only 1.931 M parameters.
Abstract
Low-light image enhancement is important for recovering usable details from photon-limited images. Existing methods often couple illumination recovery with noise amplification, and RGB-domain enhancement can introduce color shift while improving brightness. We propose the Spatial-Frequency Fusion Network (SFF-Net), which decouples intensity and chroma in HVI space and then learns frequency-aware enhancement with cross-branch compensation. The Spatial Frequency Fusion Network uses a Low-High Frequency Gated Module (LHFG) for soft spectral splitting, bounded illumination correction, and asymmetric high-frequency noise suppression, followed by a Spatial-Frequency Cross Attention Module (SFCA) that exchanges enhanced intensity cues and HV chroma features. On LOLv1, LOLv2-Real, and LOLv2-Syn, the Spatial-Frequency Fusion Network achieves PSNR/SSIM/LPIPS scores of 24.208/0.870/0.085, 23.904/0.870/0.118, and 25.713/0.941/0.045, respectively. It also improves SICE, Sony-Total-Dark, and unpaired real image results with only 1.931 M parameters.